XR-2’s 1,500-hour demonstration corpus and its DAgger correction experiment answer different questions. Demonstrations teach an initial robot policy; corrections target states that policy reaches when it goes wrong. The paper mixes correction and expert data 1:1 by trajectory count, but it does not report the total correction trajectories, frames or hours. The headline dataset size therefore cannot tell you how much correction data produced the reported improvement.
That distinction is essential when reading the results. The 84% figure comes from an expert-data scaling experiment. The rise from 58% to 93% comes from a separate checkpoint and three DAgger rounds. They are not consecutive stages of one reported run.
XR-2 is the vision-language-action model; X2W is the robot platform. DAgger means Dataset Aggregation. This review checks the September 3, 2026 paper, still at version 1, and the public artifact repository as of September 20, 2026. It does not include a dataset download, model run or robot test.
Two experiments, two baselines
The paper’s expert-data scaling section and DAgger experiment study complementary axes. Their tables and curves support comparisons within each experiment, not a synthetic 84%→93% progression.
| Experiment | Data and starting point | Author-reported result | What an adopter can conclude |
|---|---|---|---|
| Expert-data scaling | Clothes-folding subsets from 10 to 160 hours | 34% at 30 hours, 84% at 120 hours and 82% at 160 hours | More expert data improved success over the tested range before the curve flattened in this setup. The 160-hour point did not exceed the 120-hour point. |
| DAgger post-training | A separate 58% checkpoint trained on 18,695 expert trajectories across 10 subtasks | 74%, 82% and 93% after rounds one, two and three; 2.5 post-training epochs per round | Corrections improved this checkpoint over three reported rounds. The experiment does not disclose enough correction volume to convert that gain into an hours-based scaling rule. |
These are author evaluations on the paper’s tasks and system. The test settings vary material size, color and the robot’s initial position; they are not unrestricted household deployment. The result description does not supply the trial denominator needed to judge how meaningful the 84%-to-82% difference is. Neither experiment establishes the same success rates on another robot or task distribution.
The 1:1 ratio uses trajectories
The collection pipeline alternates policy execution with human takeover. An operator corrects or completes the current subtask, then returns control without restarting the whole task after each failure. The recorded segments distinguish human and policy control. This is how corrections reach states that a successful, human-led demonstration may never visit.
In the paper’s DAgger post-training method, each mixed dataset uses correction and expert data at a 1:1 ratio by trajectory count. The correction budget is weighted toward weaker subtasks. This makes the unit explicit: the ratio is not one correction hour per expert hour, and it is not a frame-level balance.
Trajectory counts can still hide large differences in duration, number of actions and task difficulty. Before copying the ratio, record the duration distribution and sampling procedure on both sides. Also record how failures select subtasks for additional collection. The authors’ 1:1 mixture is a tested protocol in this experiment, not evidence that 1:1 is universally optimal.
For a broader treatment of sampling units and oversampling effects, see the robot dataset mixture weighting guide. XR-2 adds a specific question to that workflow: whether the second source covers policy-induced failure states rather than merely adding more of the original distribution.
The paper corpus and the release card are different records
The paper reports a roughly 1,500-hour demonstration corpus: 531.7 hours of real-robot teleoperation across 32,518 trajectories and about 57.4 million frames, plus roughly 1,000 hours collected with handheld grippers using the Universal Manipulation Interface (UMI). These are demonstration counts, not the size of the correction dataset or of either experiment’s training subset.
The pinned artifact repository revision provides a separate release record. Its dated change log accounts for all 500-plus real-robot hours, 5 UMI hours and a later 100-hour UMI batch. Those entries do not by themselves verify that every file behind the paper’s 1,500-hour statement is currently downloadable. A complete file inventory or maintainer confirmation would be needed to make that claim.
The same caution applies to corrections: three collection rounds and their success rates are reported, while an exact correction quantity is not. Keep “used in the paper,” “listed on the release card” and “downloaded and validated locally” as separate fields in a reproducibility record.
A repository tree is not a reproduced training pipeline
At the pinned revision, the repository exposes directories for official and validation data, example data, Docker material and a robot description. It also includes a sample LeRobot dataset loader and an inference-interface scaffold. Those artifacts help inspect formats and integration boundaries; they do not establish that XR-2 training code or model weights are present, complete or executable.
Licensing also needs to be recorded per artifact. The repository card declares CC BY-SA 4.0, while the paper appendix and repository text describe the robot hardware description for non-commercial teaching, experimentation and academic research. The general card label does not resolve that narrower hardware-description wording. Confirm the applicable terms with the publisher before relying on commercial reuse of the URDF.
Record the missing pieces before calling it reproducible
An adopter evaluating XR-2 should capture:
- The exact paper and repository revisions, plus a file-level inventory and hashes for the data actually obtained.
- The task, subtask and trajectory-duration distributions for expert and correction data.
- The correction trigger, operator handoff, restart behavior and failure-weighted allocation rule.
- The checkpoint, mixture unit, round count, epochs per round and evaluation protocol.
- The availability and licenses of training code, configuration, weights, robot description and required dependencies.
- Results from loading samples and reproducing a bounded evaluation, kept separate from the paper’s author-reported scores.
The human-in-the-loop intervention data guide explains what to log around triggers, recovery and operator control. For XR-2, the immediate decision is narrower: do not copy the 1:1 ratio until the trajectory unit, failure distribution and correction quantity are understood, and do not call the published result reproduced until the released artifacts support the same training and evaluation path.